Pow-MUCB: A new client selection method based on Pow-d and modified UCB for federated learning in IoT

Federated learning (FL) is a collaborative machine learning (ML) approach that enables distributed training between multiple clients, achieving collective intelligence while preserving privacy. In FL, client selection (CS) plays an important role in choosing a portion of clients for training. The more efficient CS is, the more FL’s performance will be improved. A balanced CS is required to reduce the effects of discrepancies between clients’ local data and clients’ performance. For this problem, using historical data can be an effective solution. Due to a lack of reasonable and balanced use of historical data, existing CS methods in the literature suffer from serious drawbacks, like over/under-representation of clients, and data model skewness. This can adversely affect FL performance, particularly in terms of convergence rate and model accuracy. To help solve these challenges, this paper proposes a new CS method (Pow-MUCB) which is based on the Power of choice-(Pow-d-) and is equipped with a new, modified Upper Confidence Bound (UCB) approach to evaluate the clients’ contribution and performance. This method selects clients whose participation results in more balanced, representative selection and informative global updates, avoids over/ underrepresentation of clients, and model skewness, improving overall FL performance. To validate the method’s performance in both static and dynamic client sets, comprehensive comparisons and experimental results are provided. The results demonstrated that Pow-MUCB enhances the overall performance and significantly outperforms the existing baselines in terms of global model accuracy (up to 7%), convergence rate, resulting from a reduced communication rounds needed to reach the convergence.